Colorado lawmakers seek more detail on state agencies’ expanding artificial-intelligence tools
A legislative committee reviewed deployed, pilot and proposed systems for Medicaid, fraud detection, benefits processing and employment services, while agencies described safeguards and reported performance figures.
Colorado lawmakers pressed state agencies for more information about artificial-intelligence systems used for Medicaid, fraud detection, benefits processing and employment services during a Sept. 8 Joint Technology Committee hearing.
The committee reviewed a mix of deployed, pilot and proposed systems rather than a single statewide rollout. Its chair said future quarterly reports would include project and contractor information. Lawmakers also asked how the tools work, what decisions they can make and what safeguards apply, according to the committee recording.
The Colorado Department of Health Care Policy and Financing said its Medicaid fraud, waste and abuse analytics platform, called FWA Finder, generated 82 potential-investigation leads and was associated with about $200,000 in savings during the first six weeks after four providers were placed on prepayment review. HCPF said all 82 leads were classified as potential fraud, waste or abuse, but analysts, investigators and case managers must review the leads, gather documentation and decide whether further action is needed. The platform does not take enforcement action, the agency said.
FWA Finder runs more than 150 statistical models against Medicaid data, with about 75 models verified by agency analysts, HCPF officials said. The platform operates in a vendor’s Amazon Web Services cloud environment, uses multifactor authentication and National Institute of Standards and Technology controls, and is available only to HCPF fraud, waste and abuse analytics staff.
HCPF also reported that its Health First Colorado member chatbot had handled about 10,000 questions, with roughly 72% satisfaction and 89% accuracy. It answered 62% of questions without directing users elsewhere. The chatbot is limited to more than 70 approved resources and does not determine eligibility, approve or deny coverage, or make clinical decisions. Staff review member questions and generated answers, and clinical questions are referred to a nurse advice line.
The agency said users are warned not to enter protected or personal information, the system is designed to remove such information if entered, and the underlying system has a federal health privacy agreement. HCPF tested the chatbot with about 20 to 30 members before launch, officials said.
HCPF’s separate Core AI tool handles about 2,700 provider call-center interactions a month through predefined conversational flows. Officials said it does not use generative artificial intelligence or large language models and cannot approve or deny claims, change claims, determine payment amounts or make eligibility decisions. The system retains no provider or Medicaid data beyond the active transaction, though operational logs may be kept for up to 90 days for security auditing and troubleshooting.
Other systems discussed at the hearing included document-recognition technology that HCPF said processed about 200,000 benefit documents with more than 99.8% accuracy in the prior year, a limited pilot policy chatbot for state workers, and a Department of Labor and Employment fraud-screening tool that officials said reduced review time from about an hour to 15 minutes for roughly 1,000 records a month. CDLE said it was also piloting an optional artificial-intelligence mock-interview tool for job seekers and had discontinued an unemployment-insurance chatbot and interactive voice-response arrangement while seeking a single-vendor replacement.
The Colorado Bureau of Investigation was evaluating, but had not deployed, CaseGuard redaction software. CBI said the proposed local-only tool would identify more than 30 categories of sensitive information and require human review.
HCPF said its artificial-intelligence governance committee includes legal, privacy, technology, ethics, bias and program staff. New tools must undergo state technology, budget, procurement, cybersecurity and privacy reviews, officials said. The procedures fit within Colorado’s statewide artificial-intelligence guidance, which requires state risk assessments for artificial-intelligence use cases, including vendor-led projects.
Colorado’s high-risk artificial-intelligence law, Senate Bill 24-205, requires covered deployers to use reasonable care against algorithmic discrimination and establishes risk-management, impact-assessment, notice, correction and human-review requirements. The requirements began applying to high-risk systems Feb. 1, 2026.
The agencies’ reported results were presented with different measures and levels of detail. For example, HCPF did not define the savings measure or the chatbot’s accuracy denominator during the hearing, and CDLE did not describe the confirmation standard for its reported fraud-screening rate. Those differences make the figures difficult to compare across systems.